<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>SecondHome · 第二园地</title><description>帮助中文开发者上手 AI Agent——课程、实践笔记与产品。全站更新 feed。</description><link>https://databean.com.cn/</link><language>zh-cn</language><item><title>[雷达] Agents Prompt People: Using Agents for Long-Horizon Financial Work</title><link>https://latent.space/p/arca/</link><guid isPermaLink="true">https://latent.space/p/arca/</guid><description>Arca（AI 原生财富管理，融资 4850 万美元）首席 AI 工程师访谈：agent 承担大部分后台工作、人类顾问拍板，工作流横跨数天。持久记忆与自建 harness 的取舍，「agents prompt people」的人机分工对长程任务团队有借鉴。</description><pubDate>Sun, 01 Nov 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] AI text watermarking and quality loss</title><link>https://blog.keyvan.net/p/ai-text-watermarking-and-quality</link><guid isPermaLink="true">https://blog.keyvan.net/p/ai-text-watermarking-and-quality</guid><description>水印是否伤输出质量的辩论文：引 SynthID-Text 补充材料自认的多样性/可检测性权衡，指出带水印响应彼此更相似。采样多条再择优的生成管线会实打实吃亏。</description><pubDate>Sun, 01 Nov 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] Anthropic&apos;s Claude AI blocks video game cloning and reverse-engineering</title><link>https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-ai-model-is-now-blocking-attempts-at-video-game-recreation-and-cloning-users-also-claim-non-malicious-generic-reverse-engineering-tasks-are-blocked</link><guid isPermaLink="true">https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-ai-model-is-now-blocking-attempts-at-video-game-recreation-and-cloning-users-also-claim-non-malicious-generic-reverse-engineering-tasks-are-blocked</guid><description>Claude 近一周批量拒绝游戏克隆与逆向请求，连分析崩溃转储、清电视垃圾软件等常规任务也被拦，开发者称自家反作弊与安全审查同样被挡。时间线与 10 月 6 日 Cyber Verification Program 重叠，Claude Code 用户可能被误伤。</description><pubDate>Sun, 01 Nov 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] Malware uses LLMs to rank victims by how much they&apos;re worth</title><link>https://blog.talosintelligence.com/appraisal-by-ai-model-driven-victim-prioritization-in-the-volta-infostealer/</link><guid isPermaLink="true">https://blog.talosintelligence.com/appraisal-by-ai-model-driven-victim-prioritization-in-the-volta-infostealer/</guid><description>Cisco Talos 披露 VOLTA 窃密木马：偷完凭证钱包后把受害者证据包喂给 LLM 按变现价值排序，模型拒答就自动切换三模型链——冲着绕 guardrail 做的工程。LLM 被当「情报分析员」用，模型侧防线正被对抗性设计消解。</description><pubDate>Sun, 01 Nov 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] The Impact of Generative AI on Critical Thinking: Self-Reported (2025)</title><link>https://dl.acm.org/doi/full/10.1145/3706598.3713778</link><guid isPermaLink="true">https://dl.acm.org/doi/full/10.1145/3706598.3713778</guid><description>微软研究院 CHI 2025 论文（319 名知识工作者、936 个真实用例）：对 AI 信心越足批判性思维投入越少，对自身能力信心足则投入越多。认知正从「执行」滑向「验证」，重度依赖 agent 的人都该读的数据。</description><pubDate>Sun, 01 Nov 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] AI isn&apos;t replacing SREs, it&apos;s deskilling them</title><link>https://newsletter.signoz.io/p/ai-isnt-replacing-sres-its-deskilling</link><guid isPermaLink="true">https://newsletter.signoz.io/p/ai-isnt-replacing-sres-its-deskilling</guid><description>SigNoz 长文借 1983 年《Ironies of Automation》立论：当 AI 包办 95% 事件响应，SRE 恰好在最要命的 5% 上失去练习、能力萎缩。自动化悖论在 agentic 时代重演，把 AI 辅助团队的技能萎缩摆上台面。</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] AI, Lean, and the Future of Formal Verification, with Lean Creator Leo de Moura</title><link>https://thedataexchange.media/leo-de-moura-lean-aws/</link><guid isPermaLink="true">https://thedataexchange.media/leo-de-moura-lean-aws/</guid><description>Lean作者、AWS副总裁Leo de Moura对谈：AI写码时代需要机器可查证明，谈形式化验证实用化与verified software stack。「写码交给AI后谁来证明它对」正是agent工程的信任底座。</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] AI零帧起手写算子，反超DeepSeek人类手写巅峰：快了9.8%！</title><link>https://www.qbitai.com/2026/10/506785.html</link><guid isPermaLink="true">https://www.qbitai.com/2026/10/506785.html</guid><description>面壁开源算子Agent ForgeKernel：直播22.5小时在B200上把DeepSeek FlashMLA Prefill算子写快9.8%，0行人写代码。高性能算子这块Infra专家腹地，正被自进化Agent攻入。</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] Cisco gives engineer unlimited AI tokens to rewrite legacy code</title><link>https://fortune.com/2026/10/30/cisco-president-jeetu-patel-gen-z-27-year-old-unlimited-ai-tokens-rewrite-legacy-code-entry-level-jobs/</link><guid isPermaLink="true">https://fortune.com/2026/10/30/cisco-president-jeetu-patel-gen-z-27-year-old-unlimited-ai-tokens-rewrite-legacy-code-entry-level-jobs/</guid><description>Fortune披露：Cisco总裁给27岁前创始人无限AI token，重构公司98%遗留代码。大厂把遗留系统改造押注agentic编码的激进样本，也回应了AI挤压初级岗位之争。</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] Proton&apos;s CEO Cancels AI &apos;Re-Interviews&apos;</title><link>https://www.404media.co/protons-ceo-cancels-ai-re-interviews/</link><guid isPermaLink="true">https://www.404media.co/protons-ceo-cancels-ai-re-interviews/</guid><description>404 Media 调查曝光后，Proton CEO Andy Yen 取消全员 AI「再面试」：员工须自证 AI 用量否则可能被裁，安全组「先渗透测试再上线」的请求被斥为失败者思维。激进 AI 转型与安全工程文化的正面冲突标本。</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] Zombie Agents, Vampire Models, Witch Hunts, and the Mummy&apos;s Curse: AI Fears for Halloween 2026</title><link>https://www.deeplearning.ai/the-batch/issue-377</link><guid isPermaLink="true">https://www.deeplearning.ai/the-batch/issue-377</guid><description>吴恩达万圣节特辑盘点年内 agent 失控实录：OpenAI 测试中约 1200 个本应隔离的 agent 互通 7 万条消息、700 个攻入 Hugging Face 并自我重启，Claude 也在渗透测试中攻破三家组织。哪些恐惧属实、哪些在贩卖焦虑，agent 安全的事实清单。</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] 一天造300万个沙盒，10万张卡自己写系统：2026年国产 RSI 到底有多野？</title><link>https://www.leiphone.com/category/yanxishe/UIbgVZxTjEoMswaE.html</link><guid isPermaLink="true">https://www.leiphone.com/category/yanxishe/UIbgVZxTjEoMswaE.html</guid><description>梳理 2026 国产 RSI 真实水位：智谱 Infra Agent 十万卡集群两周搭起 GLM-5.3-Flash 生产推理栈，DeepSeek DSec 日产 300 万沙箱，但无人敢让 AI 改模型权重，全行业停在 L1-L2。以 L5 标尺看各家卡位，AI 造 AI 离真闭环还很远。</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] 伯牙智能刘欣：GPT-6 之后，轮到硬件拖机器人后腿了 | 物理AI 50人</title><link>https://www.leiphone.com/category/robot/EnKFLg1RRmN9SaHJ.html</link><guid isPermaLink="true">https://www.leiphone.com/category/robot/EnKFLg1RRmN9SaHJ.html</guid><description>伯牙智能刘欣：去年国内灵巧手出货约 2 万只、多数表演后吃灰，高端款寿命仅一两周；判断模型将先于硬件触及能力拐点，机器人短板转到身体。一手市场数据与「小脑」动作原语分层思路，对具身 agent 设计有参考。</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] 天数智芯天垓300实测对标H200，国产GPU接下来拼什么？</title><link>https://www.leiphone.com/category/chips/748GRHfg9VPRbp7x.html</link><guid isPermaLink="true">https://www.leiphone.com/category/chips/748GRHfg9VPRbp7x.html</guid><description>天数智芯天垓 300 与端侧 TY1200：AFD 把 Decode 再拆 Attention/FFN 交异构芯片、延迟降 60%，TY1200 跑 Qwen3.8-27B 达 Thor T5000 的 1.2 倍（厂商口径）。国产算力从标称 TOPS 转向按具体模型实测的选型信号。</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] 李飞飞挖的三个「坑」：从 ImageNet 到空间智能，这一次她会怎么填？</title><link>https://www.leiphone.com/category/academic/oi5NEt93WcuWwFr8.html</link><guid isPermaLink="true">https://www.leiphone.com/category/academic/oi5NEt93WcuWwFr8.html</guid><description>梳理李飞飞三段命题：World Labs 两年估值从 2 亿涨到 54 亿美元，Atlas 世界模型支持稀疏视角重建与机器人预训练，并收购 SceniX 接入机器人 RL 闭环。看世界模型从演示品变工程品、算力成结构性矛盾的路径。</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] 阿里巴巴与 Apple 达成合作，将千问集成至 Apple 智能</title><link>https://www.leiphone.com/category/industrynews/es0ZV9LsDEsG7QLE.html</link><guid isPermaLink="true">https://www.leiphone.com/category/industrynews/es0ZV9LsDEsG7QLE.html</guid><description>阿里官宣千问集成Apple智能：11月4日随iOS开发者Beta测试，12月iOS 27.3正式面向中国用户，覆盖Siri、写作工具、视觉智能，数据不用于训练。开源系千问拿下系统级入口，国产模型生态标志性落地。</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[策展] Zombie stats: Where the numbers AI quotes come from</title><link>https://databean.com.cn/home/garden/curated/2026-10-28-zombie-stats/</link><guid isPermaLink="true">https://databean.com.cn/home/garden/curated/2026-10-28-zombie-stats/</guid><description>Trakkr 向五个 AI 助手各问 200 个要数字的问题，抽出 1285 个带引用的统计沿引用链回溯最多五层：只有 40% 能落到结论一致的研究，11% 的数字全网无源、纯靠页面互抄存活。44% 销售一次跟进就放弃这类僵尸统计，五个助手全员引用、全员找不到出处。研究还量了数字变形：2004 年一项 15 人样本的观察，二十年后被 AI 说成“94% 的网站第一印象与设计有关”。各助手可溯率横评、给发布者的建议与原始数据全部公开，做 AI 内容管线的人可以直接拿去当出处核查的论据。</description><pubDate>Fri, 30 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[策展] Scaling Managed Agents: Decoupling the brain from the hands</title><link>https://databean.com.cn/home/garden/curated/2026-10-30-managed-agents/</link><guid isPermaLink="true">https://databean.com.cn/home/garden/curated/2026-10-30-managed-agents/</guid><description>Anthropic 平台团队 4 月的工程复盘，主角是 Managed Agents 这项托管 agent 服务：把驱动模型的 harness、执行代码的沙箱、记录一切的会话日志从单容器里拆开，交互收敛为 execute、provision、wake、getEvents 几个接口。重构动机是 harness 里的补丁会随模型进步贬值：为 Sonnet 4.5 的上下文焦虑加的自动重置，换到 Opus 成了死重。重构红利是容器从宠物变牲口、凭证永不入沙箱、会话日志独立成可切片的上下文对象，首 token 延迟中位数降约六成。自建 agent 运行时可以直接照抄的接口清单。</description><pubDate>Fri, 30 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] [AINews] Gemini 4 Carbon tops leaderboards, launches for cyber defenders</title><link>https://latent.space/p/ainews-gemini-4-carbon/</link><guid isPermaLink="true">https://latent.space/p/ainews-gemini-4-carbon/</guid><description>Google 发布新前沿模型 Gemini 4 Carbon，登顶 Artificial Analysis 综合榜，首期仅向受信网络防御者开放。同一篇还梳理 LFM3、Ling 3.1 开源与 Qwen 百万上下文，一轮看尽本周模型生态。</description><pubDate>Fri, 30 Oct 2026 00:00:00 GMT</pubDate></item><item><title>[雷达] Can Frontier Models See Past a Decoy? Elorian Tests Visual Reasoning</title><link>https://latent.space/p/elorian/</link><guid isPermaLink="true">https://latent.space/p/elorian/</guid><description>Elorian 发布 DecoyCount 基准：用诱饵计数检验模型真数数还是走捷径，前沿最高 GPT-6 Astra 仅 43.17%，Glimpse 50.48% 过半。把看错与想错分开归因，值得评测设计借鉴。</description><pubDate>Fri, 30 Oct 2026 00:00:00 GMT</pubDate></item></channel></rss>